Monimuuttujainen estimaattori teolliseen tehdassuunnitteluun
Leidén, Tuomas · Aaltodoc (Aalto University) · 2021
Paper mills are large and complex entities that transform raw material wood into finished end product by transporting and combining large masses of fibres, pulp, chemicals and water in different processes. When this kind of complex entity is engineered, the paramount thing is to be able to predict equipment quantities required for the mill and its systems as accurately as possible. Accurate estimates save costs, help optimizing resources for engineering and help to follow up the progress and readiness of the engineering phase. It also provides confidence for the client, which the project is sold to, as the estimates can be reasoned with historical data and statistics. This thesis is designed to engineering and consulting company AFRY. The motivation comes from the need of utilizing engineering data for creating better estimates of equipment counts inside a paper mill, before the project is sold to a client. The approach to the problem relies heavily on principal component analysis (PCA), that is used to analyze differences in project scopes and equipment counts. With principal component analysis combined with K-means clustering, we divide the data set into separate groups related to projects with different profiles. Accuracy of PCA is tested with explaining equipment counts with the first three principal components. A leap towards a multi-target estimator is done by mapping paper mill site principal component space and flowing substance principal component space together. Then, the location of the estimated project in site principal component space is explained by flow systems and their principal component scores. The results show that the methods used in this thesis provide a clear systematic approach for estimating the component rates of a paper mill. With clustering, we are able to divide the data into clusters that can be interpreted intuitively. With three principal components, we are able to explain over 99 % of the entire data’s variation. In further research, the similar approach can be applied for also other than paper industry, such as pulp and chemical industries.